Panel.EstimandCharacterization.OLSWeightDecomposition.Support
Per-cell measure-theoretic machinery supporting Sloczynski's weighting result in the project's regression substrate.
Basic 7 core · 8 supporting This file sets up the saturated finite-cell control class and the basic cell statistics for Słoczyński's probability-space bridge. ★ cellMean_eq_eventCondExp
Słoczyński bridge basics
This file sets up the saturated finite-cell control class and the basic cell
statistics for Słoczyński's probability-space bridge. It defines
saturatedClass, cellMass, cellShare, cellTau, propensity, and
meanReg; relates cellShare and cellTau to the shared
eventCondExp population-cell operator; and proves the elementary bounds
cellMass_nonneg, cellMass_sum_eq_one, cellShare_nonneg, and
cellShare_le_one. The membership lemmas propensity_mem_saturatedClass and
meanReg_mem_saturatedClass supply the saturated-control pieces used by the
residualization witnesses.
Linear L² control class spanned, up to almost-everywhere equality, by the finite family of cell indicators 𝟙{G = g} for g : 𝒢.
Definition (Lean source)
Cell mass (μ {G = g}).toReal, the probability weight of covariate cell g.
Definition (Lean source)
Cell-wise treatment effect E[Y(1) − Y(0) | G = g] in the shared indicator-weighted cell-mean convention. Like cellShare, zero-mass cells collapse to 0.
Definition (Lean source)
For an ambient probability space and covariate space, a probability measure μ on Ω, an integrand F, a measurable covariate map G, and a covariate cell g, the CellBridge indicator-weighted cell mean of F on {G = g} equals the shared event-level conditional expectation of F given {G = g}.
Formal statement
Proof (Lean source)
Saturated propensity propensity μ D G ω = cellShare μ D G (G ω) (pointwise, by disjointness of the {G = g} family). Plays the role of p(G(·)) and lies in saturatedClass μ G G_meas.
Definition (Lean source)
Saturated mean regression meanReg μ Y G ω is the cell-wise average of Y, evaluated at G ω. Plays the role of m(G(·)) for the outcome Y. Lies in saturatedClass μ G G_meas.
8 supporting declarations (lemmas, instances)
-
cellTau_eq_eventCondExptheorem — OLS cell treatment effect is a shared population cell mean. cellTau μ Y0 Y1 G g = E[Y(1) − Y(0) | G = g] in the shared eventCondExp operator — a genuine potential-outcome contrast.hypothesesconclusioncellTau μ Y0 Y1 G g = eventCondExp μ {ω | G ω = g} (fun ω => Y1 ω - Y0 ω)Proof (Lean source)
theorem cellTau_eq_eventCondExp {Ω 𝒢 : Type*} [MeasurableSpace Ω] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) (Y0 Y1 : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) (g : 𝒢) : cellTau μ Y0 Y1 G g = eventCondExp μ {ω | G ω = g} (fun ω => Y1 ω - Y0 ω) := cellMean_eq_eventCondExp μ (fun ω => Y1 ω - Y0 ω) G G_meas g -
cellMass_nonnegtheorem — Cell mass is nonnegative — (μ S).toReal ≥ 0.Proof (Lean source)
theorem cellMass_nonneg {Ω 𝒢 : Type*} [MeasurableSpace Ω] (μ : Measure Ω) (G : Ω → 𝒢) (g : 𝒢) : 0 ≤ cellMass μ G g := by exact ENNReal.toReal_nonneg -
cellMass_sum_eq_onetheorem — Cell masses sum to 1 under IsProbabilityMeasure μ and a measurable G valued in a finite type with MeasurableSingletonClass. The {G = g} family partitions Ω (up to μ-null sets) and the μ-mass of each is finite. This supplies the π_sum_one field for the finite-partition algebra.hypothesesΩ 𝒢 :μ :G :Ω → 𝒢G_meas :conclusion∑ g, cellMass μ G g = 1Proof (Lean source)
theorem cellMass_sum_eq_one {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsProbabilityMeasure μ] (G : Ω → 𝒢) (G_meas : Measurable G) : ∑ g, cellMass μ G g = 1 := by classical have hsum : (Finset.univ).sum (fun g => (μ (G ⁻¹' ({g} : Set 𝒢))).toReal) = (μ (G ⁻¹' (Set.univ : Set 𝒢))).toReal := by simpa [Measure.real] using (MeasureTheory.sum_measureReal_preimage_singleton (μ := μ) (s := (Finset.univ : Finset 𝒢)) (f := G) (hf := by intro g hg exact G_meas (measurableSet_singleton g)) (h := by intro g hg exact ne_of_lt <| lt_of_le_of_lt (measure_mono (Set.subset_univ _)) (by simp [IsProbabilityMeasure.measure_univ]))) simpa [cellMass, CellBridge.cellMass, preimage, preimage_univ] using hsum -
propensity_mem_saturatedClasstheorem — The pointwise representative propensity μ D G lies in saturatedClass μ G G_meas. Take the coefficient map c g := cellShare μ D G g; equality holds pointwise (and so a.e.).hypothesesΩ 𝒢 :μ :D :Ω → ℝG :Ω → 𝒢G_meas :conclusionProof (Lean source)
theorem propensity_mem_saturatedClass {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [DecidableEq 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (D : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) : (saturatedClass μ G G_meas).mem (propensity μ D G) := by exact ⟨fun g => cellShare μ D G g, Filter.EventuallyEq.rfl⟩ -
meanReg_mem_saturatedClasstheorem — The mean-regression meanReg μ Y G lies in saturatedClass μ G G_meas. Take the coefficient map c g := (∫ Y · 𝟙{G = g} dμ) / cellMass μ G g.hypothesesΩ 𝒢 :μ :Y :Ω → ℝG :Ω → 𝒢G_meas :conclusionProof (Lean source)
theorem meanReg_mem_saturatedClass {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [DecidableEq 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (Y : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) : (saturatedClass μ G G_meas).mem (meanReg μ Y G) := by refine ⟨fun g => (∫ ω', Y ω' * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω' ∂μ) / cellMass μ G g, ?_⟩ exact Filter.EventuallyEq.rfl
Integrals 1 core · 8 supporting This file proves reusable integral identities for finite covariate cells in the Słoczyński bridge. ★ integral_eq_sum_cell
Słoczyński cell integral identities
This file proves reusable integral identities for finite covariate cells in
the Słoczyński bridge. The results convert indicator-weighted integrals into
cell masses and cell means, provide integrability facts for products with
treatment and cell indicators, and reduce saturated linear-combination
orthogonality to per-cell orthogonality. The central declarations are
indicator_cell_memLp, integral_cell_indicator_one_eq_cellMass,
cell_integral_div_mul_cellMass, cellTau_mul_cellMass,
integrable_mul_indicator_D_G, propensity_eq_cellShare_of_mem,
meanReg_eq_cellMean_of_mem, integral_mul_saturated_eq_zero_of_cell, and
integral_eq_sum_cell.
For an ambient probability space and finite covariate space, a measure μ, a measurable covariate map G, and an integrable function F, the integral of F against μ equals the sum, over covariate cells g, of the integral of F restricted to the cell {G = g}.
Formal statement
Proof (Lean source)
8 supporting declarations (lemmas, instances)
-
indicator_cell_memLptheorem — Indicator helper: Set.indicator {G = g} 1 is in MemLp 2 μ for a finite measure (it is bounded by 1 and finite measure ⇒ MemLp p for every p). Stated as a separate lemma because it appears repeatedly in orthogonality and per-cell calculations.hypothesesΩ 𝒢 :μ :G :Ω → 𝒢G_meas :g :𝒢Proof (Lean source)
theorem indicator_cell_memLp {Ω 𝒢 : Type*} [MeasurableSpace Ω] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (G : Ω → 𝒢) (G_meas : Measurable G) (g : 𝒢) : MemLp (fun ω => indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω) 2 μ := by exact CellBridge.indicator_cell_memLp μ G G_meas g -
integral_cell_indicator_one_eq_cellMasstheorem — Cell indicators integrate to the corresponding real cell mass.hypothesesconclusion∫ ω, indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = cellMass μ G gProof (Lean source)
theorem integral_cell_indicator_one_eq_cellMass {Ω 𝒢 : Type*} [MeasurableSpace Ω] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) (G : Ω → 𝒢) (G_meas : Measurable G) (g : 𝒢) : ∫ ω, indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = cellMass μ G g := by simpa [cellMass] using (CellBridge.integral_cell_indicator_one_eq_cellMass μ G G_meas g) -
cell_integral_div_mul_cellMasstheorem — Dividing an indicator-weighted cell integral by a nonzero cell mass and multiplying back recovers the numerator; on zero-mass cells the numerator is zero because the indicator is a.e. zero.hypothesesProof (Lean source)
theorem cell_integral_div_mul_cellMass {Ω 𝒢 : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsFiniteMeasure μ] (F : Ω → ℝ) (G : Ω → 𝒢) (g : 𝒢) : ((∫ ω, F ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ) / cellMass μ G g) * cellMass μ G g = ∫ ω, F ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ := by simpa [cellMass] using (CellBridge.cell_integral_div_mul_cellMass μ F G g) -
cellTau_mul_cellMasstheorem — Cell-effect numerator divided by cell mass and multiplied back recovers the raw effect numerator.hypothesesProof (Lean source)
theorem cellTau_mul_cellMass {Ω 𝒢 : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsFiniteMeasure μ] (Y0 Y1 : Ω → ℝ) (G : Ω → 𝒢) (g : 𝒢) : cellTau μ Y0 Y1 G g * cellMass μ G g = ∫ ω, (Y1 ω - Y0 ω) * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ := by simpa [cellTau, cellMass] using (CellBridge.cellMean_mul_cellMass μ (fun ω => Y1 ω - Y0 ω) G g) -
integrable_mul_indicator_D_Gtheorem — Product of two singleton indicators against an L² function is integrable under a finite measure.hypothesesconclusionIntegrable (fun ω => F ω * indicator {ω' | D ω' = d} (fun _ => (1 : ℝ)) ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω) μProof (Lean source)
theorem integrable_mul_indicator_D_G {Ω 𝒢 : Type*} [MeasurableSpace Ω] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (F D : Ω → ℝ) (G : Ω → 𝒢) (D_meas : Measurable D) (G_meas : Measurable G) (F_memLp : MemLp F 2 μ) (d : ℝ) (g : 𝒢) : Integrable (fun ω => F ω * indicator {ω' | D ω' = d} (fun _ => (1 : ℝ)) ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω) μ := by let sD : Set Ω := {ω | D ω = d} let sG : Set Ω := {ω | G ω = g} have hsD : MeasurableSet sD := D_meas (measurableSet_singleton d) have hsG : MeasurableSet sG := G_meas (measurableSet_singleton g) have hmem : MemLp (fun ω => indicator sG (indicator sD F) ω) 2 μ := (F_memLp.indicator hsD).indicator hsG have hEq : (fun ω => indicator sG (indicator sD F) ω) = (fun ω => F ω * indicator sD (fun _ => (1 : ℝ)) ω * indicator sG (fun _ => (1 : ℝ)) ω) := by funext ω by_cases hG : ω ∈ sG · by_cases hD : ω ∈ sD · simp [sD, sG, indicator, hD, hG] · simp [sD, sG, indicator, hD, hG] · simp [sD, sG, indicator, hG] rw [← hEq] exact hmem.integrable (by norm_num : (1 : ENNReal) ≤ 2) -
meanReg_eq_cellMean_of_memtheorem — On its own cell, the saturated mean-regression representative equals the corresponding cell mean.hypothesesProof (Lean source)
theorem meanReg_eq_cellMean_of_mem {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] (μ : Measure Ω) (Y : Ω → ℝ) (G : Ω → 𝒢) {g : 𝒢} {ω : Ω} (hG : G ω = g) : meanReg μ Y G ω = (∫ ω', Y ω' * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω' ∂μ) / cellMass μ G g := by classical unfold meanReg rw [Finset.sum_eq_single g] · simp [hG] · intro b _ hbg simp [indicator, hG, hbg.symm] · intro hg simp at hg -
integral_mul_saturated_eq_zero_of_celltheorem — If a square-integrable residual is orthogonal to every cell indicator, it is orthogonal to every saturated finite linear combination.hypothesesΩ 𝒢 :μ :V :Ω → ℝG :Ω → 𝒢G_meas :V_memLp :MemLp V 2 μc :𝒢 → ℝhcell :∀ g, ∫ ω, V ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = 0conclusion∫ ω, V ω * (∑ g, c g * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω) ∂μ = 0Proof (Lean source)
theorem integral_mul_saturated_eq_zero_of_cell {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (V : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) (V_memLp : MemLp V 2 μ) (c : 𝒢 → ℝ) (hcell : ∀ g, ∫ ω, V ω * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = 0) : ∫ ω, V ω * (∑ g, c g * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω) ∂μ = 0 := by classical exact CellBridge.integral_mul_indicatorSpan_eq_zero_of_cell μ V G G_meas V_memLp c hcell
Orthogonality 1 core · 2 supporting This file proves the cell-level orthogonality statements needed for the Słoczyński finite-cell bridge. ★ residY_cell_orthogonal
Słoczyński orthogonality helpers
This file proves the cell-level orthogonality statements needed for the
Słoczyński finite-cell bridge. It shows that treatment and outcome residuals
are orthogonal to saturated cell indicators through
residD_cell_orthogonal and residY_cell_orthogonal. It also proves
Y_memLp_of_consistency, deriving observed-outcome square-integrability from
binary treatment, consistency, and square-integrability of the two potential
outcomes.
For an ambient probability space and finite covariate space, a square-integrable outcome Y, a measurable covariate map G, and a covariate cell g, the regression residual of Y on the saturated cell-mean model is orthogonal to the indicator of cell {G = g}.
Formal statement
Proof (Lean source)
2 supporting declarations (lemmas, instances)
-
residD_cell_orthogonaltheorem — The treatment residual is orthogonal to each saturated cell indicator.hypothesesΩ 𝒢 :μ :D :Ω → ℝG :Ω → 𝒢G_meas :D_meas :D_binary :∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1g :𝒢conclusion∫ ω, (D ω - propensity μ D G ω) * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ= 0Proof (Lean source)
theorem residD_cell_orthogonal {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (D : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) (D_meas : Measurable D) (D_binary : ∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1) (g : 𝒢) : ∫ ω, (D ω - propensity μ D G ω) * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = 0 := by classical let s : Set Ω := {ω | G ω = g} let I : Ω → ℝ := fun ω => indicator s (fun _ => (1 : ℝ)) ω let p : ℝ := cellShare μ D G g have hD_mem : MemLp D 2 μ := by have hD_bounded : ∀ᵐ ω ∂μ, D ω ∈ Icc (-1 : ℝ) 1 := by filter_upwards [D_binary] with ω hD rcases hD with hD0 | hD1 · simp [hD0] · simp [hD1] exact memLp_of_bounded (f := D) hD_bounded (D_meas.aestronglyMeasurable) (2 : ENNReal) have hI_mem : MemLp I 2 μ := by simpa [I, s] using indicator_cell_memLp μ G G_meas g have hDInt : Integrable (fun ω => D ω * I ω) μ := hD_mem.integrable_mul hI_mem have hpIInt : Integrable (fun ω => p * I ω) μ := (hI_mem.const_mul p).integrable (by norm_num : (1 : ENNReal) ≤ 2) have h_ae : (fun ω => (D ω - propensity μ D G ω) * I ω) =ᵐ[μ] (fun ω => D ω * I ω - p * I ω) := by filter_upwards [] with ω by_cases hω : ω ∈ s · have hp : propensity μ D G ω = p := propensity_eq_cellShare_of_mem μ D G hω simp [I, indicator, hω, hp, p] · simp [I, indicator, hω] have hInt : ∫ ω, (D ω - propensity μ D G ω) * I ω ∂μ = ∫ ω, D ω * I ω ∂μ - ∫ ω, p * I ω ∂μ := by calc ∫ ω, (D ω - propensity μ D G ω) * I ω ∂μ = ∫ ω, D ω * I ω - p * I ω ∂μ := integral_congr_ae h_ae _ = ∫ ω, D ω * I ω ∂μ - ∫ ω, p * I ω ∂μ := integral_sub hDInt hpIInt have hIint : ∫ ω, I ω ∂μ = cellMass μ G g := by simpa [I, s] using integral_cell_indicator_one_eq_cellMass μ G G_meas g have hpInt : ∫ ω, p * I ω ∂μ = p * cellMass μ G g := by calc ∫ ω, p * I ω ∂μ = p * ∫ ω, I ω ∂μ := integral_const_mul p I _ = p * cellMass μ G g := by rw [hIint] have hshare : p * cellMass μ G g = ∫ ω, D ω * I ω ∂μ := by simpa [p, I, s, cellShare, cellMass] using (CellBridge.cellMean_mul_cellMass μ D G g) change ∫ ω, (D ω - propensity μ D G ω) * I ω ∂μ = 0 rw [hInt, hpInt, hshare] ring -
Y_memLp_of_consistencytheorem — The observed outcome Y = D · Y1 + (1 − D) · Y0 is in MemLp 2 μ under consistency, binary treatment, and MemLp 2 μ assumptions for both potential outcomes. The pointwise bound |Y| ≤ |Y0| + |Y1| (a.e.) plus closure of MemLp 2 under sums gives the result. Stated separately because the bridge theorem takes Y_memLp as a hypothesis but downstream consumers (e.g. tests) may need to derive it.hypothesesΩ :Type*μ :Measure ΩD Y Y0 Y1 :Ω → ℝD_meas :D_binary :∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1Y0_memLp :MemLp Y0 2 μY1_memLp :MemLp Y1 2 μconsis :Y =ᵐ[μ] fun ω => D ω * Y1 ω + (1 - D ω) * Y0 ωconclusionMemLp Y 2 μProof (Lean source)
theorem Y_memLp_of_consistency {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) {D Y Y0 Y1 : Ω → ℝ} (D_meas : Measurable D) (D_binary : ∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1) (Y0_memLp : MemLp Y0 2 μ) (Y1_memLp : MemLp Y1 2 μ) (consis : Y =ᵐ[μ] fun ω => D ω * Y1 ω + (1 - D ω) * Y0 ω) : MemLp Y 2 μ := by let s1 : Set Ω := D ⁻¹' ({(1 : ℝ)} : Set ℝ) let s0 : Set Ω := D ⁻¹' ({(0 : ℝ)} : Set ℝ) have hYeq : (fun ω => indicator s1 Y1 ω + indicator s0 Y0 ω) =ᵐ[μ] (fun ω => D ω * Y1 ω + (1 - D ω) * Y0 ω) := by filter_upwards [D_binary] with ω hD rcases hD with hD0 | hD1 · simp [s1, s0, hD0] · simp [s1, s0, hD1] have h_mem : MemLp (fun ω => indicator s1 Y1 ω + indicator s0 Y0 ω) 2 μ := by have hD1_meas : MeasurableSet s1 := by simpa [s1] using D_meas (measurableSet_singleton (1 : ℝ)) have hD0_meas : MeasurableSet s0 := by simpa [s0] using D_meas (measurableSet_singleton (0 : ℝ)) exact (Y1_memLp.indicator hD1_meas).add (Y0_memLp.indicator hD0_meas) have h_mem' : MemLp (fun ω => D ω * Y1 ω + (1 - D ω) * Y0 ω) 2 μ := by exact (memLp_congr_ae hYeq).1 h_mem exact (memLp_congr_ae consis).2 h_mem'
Partition 4 core · 1 supporting This file constructs the finite partition and residualization witnesses used to connect saturated-control population objects with Słoczyński's finite-partition OLS weight algebra. ★ partitionOf_p_eq_eventCondExp
Słoczyński partition bridge
This file constructs the finite partition and residualization witnesses used to
connect saturated-control population objects with Słoczyński's finite-partition
OLS weight algebra. The definition partitionOf packages measurable binary
treatment and a finite cell classifier into the FinitePartition consumed by
the algebraic theorems, while partitionOf_p_eq_eventCondExp and
partitionOf_tau_eq_eventCondExp certify that its treated shares and
treatment effects are shared population cell means. The definitions
residWitnessD and residWitnessY build the residualization witnesses for the
treatment and observed outcome variables against the saturated control class.
Measurable binary treatment data with a finite cell classifier and positive saturated-overlap denominator determine the finite Słoczyński partition of cell masses, treated shares, and within-cell treatment effects.
Definition (Lean source)
Population-cell certificate for partitionOf. For a probability space (Ω, μ) with a measurable finite covariate G and treatment/potential-outcome data D, Y0, Y1, suppose treatment is binary almost everywhere and the covariate cells have nondegenerate treatment overlap. Then, for any cell g, the treated share p_g of the Słoczyński partition partitionOf equals the shared population cell mean E[D ∣ G = g] (via eventCondExp).
Formal statement
Proof (Lean source)
Residualization witness for the treatment variable D.
Definition (Lean source)
Residualization witness for the outcome variable Y.
Definition (Lean source)
1 supporting declaration (lemmas, instances)
-
partitionOf_tau_eq_eventCondExptheorem — Population-cell certificate for partitionOf. The cell treatment effect τ_g of the Słoczyński partition is the shared population potential-outcome contrast E[Y(1) − Y(0) | G = g] (via eventCondExp), so the overlap-weighted estimand is built from genuine potential-outcome cell means.hypothesesΩ 𝒢 :μ :D Y0 Y1 :Ω → ℝG :Ω → 𝒢G_meas :D_binary :∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1g :𝒢conclusion(partitionOf μ D Y0 Y1 G G_meas D_binary overlap).τ g= eventCondExp μ {ω | G ω = g} (fun ω => Y1 ω - Y0 ω)Proof (Lean source)
theorem partitionOf_tau_eq_eventCondExp {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [DecidableEq 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsProbabilityMeasure μ] (D Y0 Y1 : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) (D_binary : ∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1) (overlap : 0 < ∑ g, cellMass μ G g * (cellShare μ D G g * (1 - cellShare μ D G g))) (g : 𝒢) : (partitionOf μ D Y0 Y1 G G_meas D_binary overlap).τ g = eventCondExp μ {ω | G ω = g} (fun ω => Y1 ω - Y0 ω) := cellTau_eq_eventCondExp μ Y0 Y1 G G_meas g
PerCell 1 core · 1 supporting This file proves the per-cell denominator and numerator identities used in the Słoczyński finite-cell bridge. ★ num_per_cell
Słoczyński per-cell bridge identities
This file proves the per-cell denominator and numerator identities used in
the Słoczyński finite-cell bridge. These identities reduce the
Frisch-Waugh-Lovell denominator and numerator to cell masses, treated shares,
and cell-specific treatment effects before summing over the finite partition.
The theorem denom_per_cell identifies the residualized-treatment second
moment on one cell, and num_per_cell identifies the residualized
treatment-outcome covariance on one cell under consistency and the finite-cell
conditional-mean-independence bridge condition.
Per-cell numerator identity. For a probability space (Ω, μ) with measurable treatment D, an outcome Y, square-integrable potential outcomes Y0, Y1, and a measurable finite covariate G, suppose treatment is binary almost everywhere, the observed outcome is consistent — Y equals D·Y1 + (1−D)·Y0 almost everywhere, and the finite-cell conditional-mean-independence bridge condition holds, the integrated finite-cell substitute for E[Y(d) ∣ D, G] = E[Y(d) ∣ G]. Then, for any cell g, the propensity-residual-weighted, cell-indicator-integrated outcome ∫ (D − propensity) · Y · 𝟙{G = g} dμ equals cellMass g · cellShare g · (1 − cellShare g) · cellTau g.
Formal statement
Proof (Lean source)
1 supporting declaration (lemmas, instances)
-
denom_per_celltheorem — Per-cell denominator identity. On the cell {G = g}, (D − propensity)(ω) = D ω − cellShare μ D G g; squaring and using D ∈ {0, 1} plus the defining identity of cellShare giveshypothesesΩ 𝒢 :μ :D :Ω → ℝG :Ω → 𝒢G_meas :D_meas :D_binary :∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1g :𝒢conclusion∫ ω, (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μProof (Lean source)
theorem denom_per_cell {Ω 𝒢 : Type*} [MeasurableSpace Ω] [Fintype 𝒢] [DecidableEq 𝒢] [MeasurableSpace 𝒢] [MeasurableSingletonClass 𝒢] (μ : Measure Ω) [IsFiniteMeasure μ] (D : Ω → ℝ) (G : Ω → 𝒢) (G_meas : Measurable G) (D_meas : Measurable D) (D_binary : ∀ᵐ ω ∂μ, D ω = 0 ∨ D ω = 1) (g : 𝒢) : ∫ ω, (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * indicator {ω' | G ω' = g} (fun _ => (1 : ℝ)) ω ∂μ = cellMass μ G g * (cellShare μ D G g * (1 - cellShare μ D G g)) := by let s : Set Ω := {ω | G ω = g} let I : Ω → ℝ := fun ω => indicator s (fun _ => (1 : ℝ)) ω let p : ℝ := cellShare μ D G g let A : ℝ := ∫ ω, D ω * I ω ∂μ let M : ℝ := cellMass μ G g have hD_mem : MemLp D 2 μ := by have hD_bounded : ∀ᵐ ω ∂μ, D ω ∈ Icc (-1 : ℝ) 1 := by filter_upwards [D_binary] with ω hD rcases hD with hD0 | hD1 · simp [hD0] · simp [hD1] exact memLp_of_bounded (f := D) hD_bounded (D_meas.aestronglyMeasurable) (2 : ENNReal) have hI_mem : MemLp I 2 μ := by simpa [I, s] using indicator_cell_memLp μ G G_meas g have hDInt : Integrable (fun ω => D ω * I ω) μ := hD_mem.integrable_mul hI_mem have hconstIInt : Integrable (fun ω => (p * p) * I ω) μ := (hI_mem.const_mul (p * p)).integrable (by norm_num : (1 : ENNReal) ≤ 2) have hlinInt : Integrable (fun ω => (1 - 2 * p) * (D ω * I ω)) μ := hDInt.const_mul (1 - 2 * p) have h_ae : (fun ω => (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * I ω) =ᵐ[μ] (fun ω => (1 - 2 * p) * (D ω * I ω) + (p * p) * I ω) := by filter_upwards [D_binary] with ω hD by_cases hω : ω ∈ s · have hp : propensity μ D G ω = p := propensity_eq_cellShare_of_mem μ D G hω rcases hD with hD0 | hD1 · simp [I, indicator, hω, hp, p, hD0] · simp [I, indicator, hω, hp, p, hD1] ring · simp [I, indicator, hω] have hmain : ∫ ω, (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * I ω ∂μ = (1 - 2 * p) * A + (p * p) * M := by calc ∫ ω, (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * I ω ∂μ = ∫ ω, (1 - 2 * p) * (D ω * I ω) + (p * p) * I ω ∂μ := integral_congr_ae h_ae _ = ∫ ω, (1 - 2 * p) * (D ω * I ω) ∂μ + ∫ ω, (p * p) * I ω ∂μ := integral_add hlinInt hconstIInt _ = (1 - 2 * p) * A + (p * p) * M := by rw [integral_const_mul, integral_const_mul] have hIint : ∫ ω, I ω ∂μ = M := by simpa [I, s, M] using integral_cell_indicator_one_eq_cellMass μ G G_meas g simp [A, M, hIint] have hshare : p * M = A := by simpa [p, M, A, I, s, cellShare, cellMass] using (CellBridge.cellMean_mul_cellMass μ D G g) change ∫ ω, (D ω - propensity μ D G ω) * (D ω - propensity μ D G ω) * I ω ∂μ = M * (p * (1 - p)) rw [hmain] rw [← hshare] ring